Deep Learning Prediction of Voxel-Level Liver Stiffness in Patients with Nonalcoholic Fatty Liver Disease
Companion software for https://pubs.rsna.org/doi/abs/10.1148/ryai.2021200274
NOTE: The images used for training and evaluation are private data and NOT included in this repo.
This repo contains notebooks and python modules used to train and evaluate the MRE AI architecture. All input images are assumed to be properly oriented and converted into NIFTI format. This repo covers registration, segmentation, data container generation, model training and evaluation.
All required packages can be installed using the environment.yml file:
conda env create -f environment.yml
Registration is handled via SimpleElastix, which is an extension of SimpleITK. It can be built via the instructions found here: https://simpleelastix.readthedocs.io/GettingStarted.html. Please do not install any other versions of SimpleITK, as this can conflict with SimpleElastix.
It is recommended that the mre_ai modules be installed via pip:
pip install -e .
All relevant python modules are found in the mre_ai/ directory. Segmentation is handled by train_seg_model.py and either pytorch_arch_deeplab_3d.py, as was done in the paper, or pytorch_arch_models_genesis.py (ModelsGenesis github which can show improved results and generalization. MRE generation is handled by train_mre_model.py and pytorch_arch_deeplab_3d.py.
Please refer to notebooks for information on how registration, segmentation, training, and evaluation are conducted.